Roll a wheel along a curve and its trip-counter reads the arc length — the exact value of the integral \( \int_a^b \sqrt{1+f'(x)^2}\,dx \). Every tiny turn of the wheel adds a hypotenuse \( ds=\sqrt{dx^2+dy^2} \): flat stretches barely move the dial, steep climbs spin it fast. Drag the curve into hills and valleys and watch the odometer update live.
Zoom in on any smooth curve until a tiny piece of it looks straight. Over a horizontal step \( dx \) the curve also rises by \( dy=f'(x)\,dx \), so the actual length of that little diagonal piece — the distance the wheel truly rolls — is the hypotenuse of a right triangle:
\( ds=\sqrt{dx^2+dy^2}=\sqrt{1+\left(\tfrac{dy}{dx}\right)^2}\;dx=\sqrt{1+f'(x)^2}\;dx. \)
The odometer never measures the curve all at once. It just keeps adding these micro-hypotenuses as it turns. Adding up (integrating) every \( ds \) from \( x=a \) to \( x=b \) gives the total distance travelled — the arc length:
\( L=\int_a^b ds=\int_a^b \sqrt{1+f'(x)^2}\;dx. \)
The factor \( \sqrt{1+f'^2} \) is the local stretch: where the road is flat, \( f'=0 \), the factor is 1 and the wheel gains one unit of trip per unit of \( x \) — the dial matches the horizontal span. Where the road is steep, \( f' \) is large and the factor blows up, so a small advance in \( x \) still spins the wheel a long way. This is why the total \( L \) is always at least the straight span \( b-a \), with equality only for a horizontal line. In the applet the green ds strips are a live Riemann sum: each strip's height is \( \sqrt{1+f'^2} \) at its midpoint, its area approximates the arc contributed by that slice, and the strips' total tracks the odometer's reading.
Two honest subtleties worth noticing while you drag. First, arc length is an integral of a derivative: it needs the slope \( f' \) everywhere, which is why a curve with a sharp corner (try the Spike) still has finite length but its integrand \( \sqrt{1+f'^2} \) spikes where the slope jumps. Second, the number the applet prints is a genuine numerical integral — the curve is sampled into several hundred short segments and their true lengths \( \sqrt{\Delta x^2+\Delta y^2} \) are summed, so the odometer and the printed \( L \) agree to the pixel. This same "sum of little hypotenuses" idea generalizes directly: a parametric curve uses \( ds=\sqrt{\dot x^2+\dot y^2}\,dt \), and a path in space just adds a \( \dot z^2 \) under the root.
Further reading: any single-variable calculus text on "arc length" (e.g. Stewart, Calculus, §8.1); for the parametric and vector versions see the line-integral chapter. The curve here is a Catmull–Rom spline through the six draggable handles, so it stays a smooth function of \(x\) as you reshape it.